Welcome to the main channel on the development of MoarVM, a virtual machine for NQP and Rakudo (moarvm.org). This channel is being logged for historical purposes.
Set by lizmat on 24 May 2021.
00:02 reportable6 left 00:05 reportable6 joined 01:28 frost joined 02:03 nine left, nine joined 04:54 squashable6 left 04:57 squashable6 joined 05:06 squashable6 left
Nicholas good *, #moarvm 05:44
05:49 kjp joined
japhb Good *, Nicholas 05:57
06:03 reportable6 left 06:05 reportable6 joined
nine Better *, all 06:11
08:08 squashable6 joined 09:00 japhb left 09:12 japhb joined
jnthnwrthngtn moarning o/ 09:21
Nicholas \o
nine jnthnwrthngtn: in github.com/MoarVM/MoarVM/commit/8b...b8ffaebdbb you added the deopt pre ins and deopt synth annotations. Presumably the deopt pre ins annotation is there to contain the correct deopt destination PC. But the deopt synth annotation contains the index that the materializations are registered for. 10:04
In the failing example, all stacked guards deopt to the same: -> pc 218 10:09
But I don't know if that's coincidence or if that will always be true. In the latter case I wonder what the pre deopt ins annotation is needed for 10:10
jnthnwrthngtn Any guards stacked up ahead of a runbytecode, be they a product of dispatch program translation or of ensuring we can link a specialization or inline, should deopt to immediately before the dispatch op 10:14
Which is why it's a pre-deopt point
Which is why it's a pre-deopt point 10:15
We also dispatch_o is a curious case:
dispatch_o .d w(obj) str callsite :cache :deoptallpoint :predeoptonepoint :maycausedeopt :specializable :postlogged :deoptonepoint
It actually has every kind of deopt point.
nine This looks like the synt points already will use the index of a pre-deopt point: ann->data.deopt_idx = predeopt_index; So why then add an additional deopt-pre-ins annotation?
jnthnwrthngtn 1. predeopt because if the guards stacked up ahead fail we need to deopt to before the dispatch instruction
1. predeopt because if the guards stacked up ahead fail we need to deopt to before the dispatch instruction 10:16
2. deoptonepoint for if we inserted a return value guard afterwards
3. deoptallpoint for global deopt
nine: Oh, I see what you mean...totally missed that it was doing that when I first read 10:18
Maybe that's not needed, given we just add the synth one in src/spesh/disp.c 10:19
10:19 brrt joined
Nicholas good *, brrt 10:22
brrt \o Nicholas 10:24
does jnthn have is normal name again
not yet...
MasterDuke timo: github.com/MoarVM/MoarVM/blame/mas...ion.c#L960 might be something for you 10:55
jnthnwrthngtn Yay, enough megamorphic handlers done that the rather involved Agrammon application now never fills up a dispatch cache. 11:09
Well, on the use case I threw at it, anyway 11:11
11:26 brrt left 11:36 brrt joined
Geth MoarVM: MasterDuke17++ created pull request #1571:
Fix some of the things found by the clang static analyzer
11:59
12:02 reportable6 left
Geth MoarVM: d12d6f35cb | (Daniel Green)++ | src/6model/reprs/MVMCapture.c
Add some context to capture arg exceptions

Say what op they were thrown from (if any) and include both the given and allowed values.
12:31
MoarVM: 7b5b6bad77 | MasterDuke17++ (committed using GitHub Web editor) | src/6model/reprs/MVMCapture.c
Merge pull request #1566 from MasterDuke17/add_some_context_to_capture_arg_exceptions
MoarVM: b3e14b1512 | (Daniel Green)++ | src/profiler/configuration.c
Add default case when determining an operand size

Otherwise GCC at the -O2 optimization level warns that `size` may be used uninitialized.
12:32
MoarVM: 4293ea853a | (Daniel Green)++ | src/strings/ops.c
Silence a GCC warning at optimization level -O2...

by twice moving a conditional in MVM_string_index outside a loop, so it won't potentially be used uninitialized. I think it couldn't actually happen, but this is probably a little faster also.
MoarVM: d48a5a4129 | MasterDuke17++ (committed using GitHub Web editor) | 2 files
Merge pull request #1570 from MasterDuke17/fix_warnings_found_by_gcc_at_O2_optimization_level
MoarVM/master: 4 commits pushed by (Daniel Green)++, MasterDuke17++ 12:36
13:03 reportable6 joined 14:03 linkable6 left, evalable6 left 14:04 evalable6 joined 14:16 frost left 15:05 linkable6 joined 15:09 brrt left
timo the moarvm talk just ended half an hour ago? 15:32
MasterDuke did anyone here register to listen/attend live? some of the other talks looked interesting too, but i couldn't quite justify the cost 15:37
btw, the clang static analyzer found a lot more things than those couple i fixed, i just took the ones that were really easy to verify 15:38
jnthnwrthngtn I survived the talk. And some questions afterwards. :) 15:44
It was recorded and I believe the video will be out later today 15:47
MasterDuke nice 15:53
if anyone wants to repro, it was really easy, i just did a configure and make clean, then `scan-build make -j12` 15:58
sena_kun jnthnwrthngtn++ 16:00
timo later today sounds great 16:12
16:15 patrickb joined 16:17 patrickb left 16:59 vrurg left 17:04 vrurg joined
[Coke] jnthnwrthngtn: doing the work, talking the talk... 17:28
17:43 linkable6 left 18:02 reportable6 left 18:05 reportable6 joined
timo sliding the slides 18:32
Nicholas drinking the tea 18:39
aha, but more
punning the puns
I believe that's more important round here. 18:40
lizmat was punning in court for 4 hours today 18:46
apart from being in the car for 5+ hours today 18:47
lizmat is very tired and calls it an early day 18:48
Nicholas that doesn't sound that fun
lizmat no, it wasn't, but at least it was hopefully the last time in this case
&
timo casing the case 18:51
[Coke] thinks the last time he was in court, he punned a juror. 18:58
Nicholas so where on bittorrent is this talk video then? :-) 19:00
timo slaps th emule 19:02
Nicholas I suspect that at some point we might find it here www.youtube.com/channel/UCwG9512Wm...6Iqshz4Dpg 19:04
MasterDuke "now that's a name i haven't heard in a long time" 19:06
timo hope someone set up a camera on a tripod in an empty theater :P 19:08
gotta see the newest moarvlm flick as soon as it comes out
MasterDuke timo: btw, why is it that some small scripts create gigantic profiles? is that just a limitation of how we collect profile data? or is it indicative of something non-optimal happening when rakudo/moarvm run the code? 19:12
[Coke] Nicholas linked to the channel, is this one the talk? youtu.be/dQw4w9WgXcQ 19:13
Nicholas :-)
score!
timo MasterDuke: probably all my fault :) 19:36
MasterDuke i guess that calls for 3 smacks with a wet noodle, as my grandmother would say 19:37
timo cruel o_O
MasterDuke i guess it depends on whether they're ramen or udon 19:38
timo canneloni perhaps
MasterDuke for example, on a branch of Text::Diff::Sift4 that reverts a bunch of converting stuff to nqp ops, profiling `use Text::Diff::Sift4; my $a; for "words_1k.txt".IO.lines -> $word { $a := sift4($word, "testing") }; say now - INIT now` creates something like a 16mb sql file 19:42
timo i think we still have that thing where we accidentally record recursions instead of loops under certain circumstances involving inlining 19:43
MasterDuke while on master (which doesn't change any overall structure, just uses nqp ops in a bunch of places, using a 10k file only creates a 578k html file
hm, that sounds a little familiar. i don't remember if there was a proposed fix though? 19:44
the sift4 code doesn't have any explicit recursion 19:45
timo right
MasterDuke ha! 16mb on that branch normally. 330k on that branch with MVM_SPESH_INLINE_DISABLE=1 19:47
and runtime is only 0.27s 19:48
19:53 evalable6 left
timo i might actually put in a "don't do this optimization if profiling is turned on" so we get correct profiles until i fix it properly 19:54
MasterDuke well, inlining is so good for performance it might significantly throw off total runtime measurements 19:59
do you know how to fix it? 20:00
timo i've got to hope :P 20:06
not every inlining causes the problem, only some specific cases 20:08
so i would only disable some inlinings, or even better, some secondary optimization
20:46 linkable6 joined
jnthnwrthngtn I wonder if the optimization where we rewrite exception handling to goto (which can happen over inline boundaries) fails to correctly handle leaving of the inline 20:47
MasterDuke the kind of exception handling that's used to implement `last`? because i do use that in the code 20:49
timo i believe that's the one, yeah, jnthnwrthngtn 20:50
jnthnwrthngtn MasterDuke: It may manage that, yes 20:51
timo did you continue the code that puts info about inline caches / morphism-amount in the profiler output?
jnthnwrthngtn I think it does control exceptions mostly (that doesn't have an exception object)
timo: I didn't get it to output the SQL yet 20:52
timo: But the JSON makes it into the HTML output, so "just" needs me (or somebody else!) to remember how to JavaScript + Angular
Geth MoarVM: MasterDuke17++ created pull request #1572:
Lego jit implementation of nqp::abs_i
timo oh, is it pushed up already? 20:53
jnthnwrthngtn The MoarVM bit is, and I think it automatically makes it into the JSON blob as a result
627c92ccb00 in MoarVM 20:54
timo cool, i'll have a look while waiting for your talk's VOD or something :)
20:54 linkable6 left
jnthnwrthngtn The profiler UI is the only thing I ever wrote in Angular; the next web app I needed to make I did in React, and then I decided I vastly preferred it. :) 20:55
(Preferred React, to be clear.) 20:56
afk for a bit
timo yeah, react somehow makes more sense to me as well. perhaps a lot of that is how angular split things into components and modules or something like that? didn't get my head wrapped around that at that point, but react is a lot more explicit about where things come from 21:03
japhb jnthnwrthngtn: Well, to be fair, the creators of AngularJS seemingly preferred something else as well (since Google went "Dear me, don't use AngularJS, use Angular2 instead -- it's a complete ground-up rewrite! Everything's different! We swear it's better now!") 21:19
Even within Google, those of us on the "consumer" side of those frameworks looked askance across the campus at the "producer" side. 21:20
I think my personal reaction was something like "Fool me once ..." 21:22
timo "askance"?
do we have vscode users in here? i can't seem to find a list of commits in the VCS panel? 21:23
japhb www.dictionary.com/browse/askance 21:24
(Not being a jerk -- it's actually a pretty clear definition)
timo i see
japhb wonders if the suspicion variant was based on the sideways look variant -- because you wanted to at least keep things you couldn't trust in your peripheral vision, and subtly check on them every so often 21:26
timo jnthnwrthngtn: i thought we wanted the info about mono/poly/mega in a separate place that's per-routine rather than per-callgraph-node 21:52
21:55 linkable6 joined
timo no recording yet, huh? 23:36
23:55 evalable6 joined